Better timeslices

Finally fixes #7 and produces a much nicer looking fit for large tree sequences, but leads to slightly worse performance for tiny tree sequences such as those tested in test_accuracy.py, because of #230. When we fix that, this PR should provide uniformly better performance, I hope

Author: hyanwong

CHANGELOG.md

@@ -11,6 +11,9 @@
 - The default `min_branch_length` and `eps` have been set to 1e-8 rather than 1e-10, to avoid
   occasional issues with floating point error.
 
+- A slight adjustement in the default timeslices used for the discrete-time method: see
+  https://github.com/tskit-dev/tsdate/pull/233 for a rationale.
+
 ## [0.2.4] - 2025-09-18
 
 - Add support for Python 3.13, minimum version is now 3.10.

tsdate/prior.py

@@ -936,8 +936,10 @@ def create_timepoints(base_priors, n_points=21):
     # missing samples, otherwise we only have one set of priors anyway
     prior_params = base_priors.prior_with_max_total_tips()
     # Percentages - current day samples should be at time 0, so we omit this
-    # We can't include the top end point, as this leads to NaNs
-    percentiles = np.linspace(0, 1, n_points + 1)[1:-1]
+    # We can't include 1 at the top end as it will be at infinite time
+    # so we take the upper time point to be a fraction lower than 1 (as if we divided
+    # the last timeslice into n_points evenly picked quantiles and removed the last one)
+    percentiles = np.linspace(0, 1 - 1 / (n_points**2), n_points)[1:]
     # percentiles = np.append(percentiles, 0.999999)
     param_cols = np.where([f not in ("mean", "var") for f in PriorParams._fields])[0]
     """